In Claude 3.5, Anthropic introduced the term "artifact" so that they could instruction-fine-tune the model on the term, which is different than common terms like "asset" or "attachment", and not very…

1 year, 5 months

Basically, I think the UX of ChatGPT (both web & mobile app) has its flaws but is by far the best out of end-user LLM offerings I've tried (at the paid tier): ChatGPT, Google Gemini, You.com, Claude.…

1 year, 9 months

However, only the ChatGPT service (website & iOS app) let me paste pretty long inputs into the text box. Poe & You.com tell me that the prompt is too long and tell me to "upload a TXT file" instead, …

1 year, 9 months

I AM really impressed with You.com Pro which is $15/month, has unlimited access to both GPT-4 and GPT-4 Turbo (you can explicitly choose), also to Claude 2 & Gemini Pro, and it has several websearch-…

1 year, 9 months

The limitation of 4000 chars in Copilot makes many prompts unfeassble. On the other hand, Copilot Pro gives me access to "GPT-4 Creative" (the REAL GPT-4), "GPT-4 Turbo Creative", and "GPT-4 Precise"…

1 year, 9 months

My MS Copilot Pro iOS app gives me just 4000 characters input, which is not great. On the other hand, its web search integration seems much better than that of ChatGPT 4, but worse than You.com Pro e…

1 year, 9 months

Groq has an API waitlist at https://console.groq.com/ — its speed is truly insane, especially with Mixtral 4x7b. Mixtral is a very able model and I think if served via Groq, it can truly serve as a…

1 year, 9 months

> https://thisdayinai.com/post/105-groqs-language-processing-unit-lputm-is-worlds-fastest-llm-speed-alpha-demo/ Apart grom the Groq website, these models are available through the Poe.com subscripti…

1 year, 9 months

> https://thisdayinai.com/post/104-gemini-15-outshines-gpt-4-turbo-128k-on-long-code-prompts-hvm-author/ This is comparing Gemini Pro 1.5 MoE with GPT-4 Turbo, which is a fair comparison. GPT-4 Turb…

1 year, 9 months

Wrong thread

1 year, 9 months

This is comparing Gemini Pro 1.5 MoE with GPT-4 Turbo, which is a fair comparison. GPT-4 Turbo isn't GPT-4, and Gemini Pro isn't Gemini Ultra. However, I have a strong feeling that Google has worke…

1 year, 9 months

People have demonstrated that LLMs with 32k-100k input contexts can (with the right prompting) summarize the "narrative flow" of a long text. If they can summarize, then they can also do things like …

1 year, 9 months

Obviously the same is true for all kinds of "consistent text corpus" work: the StarTrek creators can stuff the entire collection of scripts for all the StarTrek movies and shows, and have the model s…

1 year, 9 months

A context window of 1M tokens or even 10M tokens is extremely useful for literature creation. Let's say anauthor has 5 novels she has written, and she wants to use an LLM as a writing aid for the 6th…

1 year, 9 months

I cannot find it anymore but after Google introduced the Gemini "Advanced" chatbot plan, someone from Google said that for some operations like image analysis, the chatbot still uses the Gemini Pro m…

1 year, 10 months

I've been using "ChatGPT 4" (= GPT-4 Turbo 128K/4K with automatic tool switching) vs. "ChatGPT Classic" (= GPT-4 8K) for coding in ChatGPT for the last months, and the Turbo model has always been "la…

1 year, 10 months

OpenAI made this all very confusing. Many people think that the GPT-4 Turbo models are supposed to be an "improvement" over GPT-4: no, they're not, they're a different line of models: faster but much…

1 year, 10 months

The "GPT 4 Turbo" models are nothing like the older "GPT 4" models: they're very obviously distilled and/or quantized. Also, Turbo have 128K input context but a max of 4K output. The classic GPT-4 (a…

1 year, 10 months